Skip to content

coding-agents.md

Latest commit

 

History

History
39 lines (31 loc) · 1.86 KB

coding-agents.md

File metadata and controls

39 lines (31 loc) · 1.86 KB
Raw

Use coding agents

You can send coding-agent traffic, such as Claude Code, through DeltaLLM. DeltaLLM then applies keys, budgets, rate limits, guardrails, and usage reports to each developer.

Claude Code

Claude Code uses the Anthropic Messages API. Use a model that has a deployment with provider: anthropic. DeltaLLM then sends each request to Anthropic unchanged, so extended thinking, prompt caching, images, and beta features work. See Anthropic Messages.

  1. Create an application key for the developer. Set a budget and an expiry date on the key. See Application keys.
  2. Give the key access to the Claude models that the developer can use.
  3. Set these environment variables for the developer:
export ANTHROPIC_BASE_URL="https://your-gateway.example.com"
export ANTHROPIC_API_KEY="YOUR_DELTALLM_KEY"
# Use your DeltaLLM model names.
export ANTHROPIC_MODEL="claude-sonnet-4-6"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="claude-haiku-4-5"
  1. Start claude. Each request uses the DeltaLLM key.

Notes:

  • Claude Code can read the model list from GET /v1/models. Set CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1 to turn this on. Claude Code shows only model names that contain claude or anthropic.
  • Set the model price input_cost_per_token_cache_hit. Claude Code reads most of its prompt from the Anthropic prompt cache, and DeltaLLM bills cache reads at this price.
  • If the model also has deployments with other providers, those deployments use the translation hop. On those deployments, thinking and prompt caching are not available.
  • Claude Code counts tokens with POST /v1/messages/count_tokens. DeltaLLM asks Anthropic for the count, or returns an estimate for other providers. DeltaLLM does not bill these calls. See Count tokens.